Thinking Machines Lab, founded by Mira Murati, has unveiled its first product — the “Tinker” API, designed for fine-tuning language models, including large models with “multi-expert” architecture.
Currently, the beta version of Tinker is being tested in closed mode among researchers and developers, with a waiting list available for those interested in accessing it. The API runs on the Python programming language.
The service currently supports various versions of Qwen 3 from China’s Alibaba, as well as Llama 3.1, 3.2, and 3.3 from Meta. All supported models are open-source, and the list is expected to expand in the future.
The fine-tuning process is carried out using the LoRA (Low-Rank Adaptation) method. Instead of modifying all parameters of the base language model, LoRA makes compact additions to the existing architecture, thereby saving computational resources.
To use the Tinker API, users need to provide a sample dataset (recommended size not specified). The fine-tuning process itself is performed directly on Thinking Machines Lab’s AI infrastructure. The company allocates the resources and monitors the system to resolve any potential failures. At the end of the process, users can download the weights of the fine-tuned model.
Alongside the API, the startup has also released a library called Tinker Cookbook, which includes ready-to-use algorithms built on top of the system.
It should be noted that the Tinker service is initially offered free of charge, but the startup has announced that it will transition to a paid system in the “coming weeks,” depending on usage volume. In July 2025, Thinking Machines Lab raised $2 billion in funding, bringing its total valuation to $12 billion.
